Home    中文  
 
  • Search
  • lucene Search
  • Citation
  • Fig/Tab
  • Adv Search
Just Accepted  |  Current Issue  |  Archive  |  Featured Articles  |  Most Read  |  Most Download  |  Most Cited

Chinese Journal of Cerebrovascular Diseases(Electronic Edition) ›› 2026, Vol. 20 ›› Issue (03): 308-319. doi: 10.3877/cma.j.issn.1673-9248.2026.03.011

• Evidence Based Medicine • Previous Articles    

Diagnostic performance of different artificial intelligence models for poor stroke outcome

Yifan Liang1, Jingyu Mou1, Yating Wu1, Jingtao Pi1, Le Chen1, Jian Wu1,2,3,()   

  1. 1 Department of Neurology, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua Medicine, Tsinghua University, Beijing 102218, China
    2 School of Healthcare Management, Tsinghua University, Beijing 100084, China
    3 IDG/McGovern Institute for Brain Research at Tsinghua University, Beijing 100084, China
  • Received:2026-02-02 Online:2026-06-01 Published:2026-06-23
  • Contact: Jian Wu

Abstract:

Objective

To systematically evaluate and compare the diagnostic performance of different artificial intelligence (AI) models in predicting poor functional outcomes in patients with acute ischemic stroke, with a particular focus on the trade-off between sensitivity and specificity across models.

Methods

PubMed, Web of Science, Embase, The Cochrane Library, China National Knowledge Infrastructure, and Wanfang databases were systematically searched to identify studies applying AI-based models for functional outcome prediction after stroke. According to modeling principles, the models were categorized into regression models, single decision trees, random forest (RF) models, boosting ensemble models, support vector machines (SVM), deep learning (DL) models, and other machine learning approaches. The 2×2 contingency table data were extracted or calculated. Diagnostic test accuracy Meta-analysis was performed to pool sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, diagnostic odds ratio, and the area under the receiver operating characteristic curve (AUROC). Summary receiver operating characteristic (SROC) curves were used to compare the overall diagnostic performance of different model categories.

Results

A total of 32 studies involving 56 458 patients were included. The pooled AUROC values across different model categories ranged from 0.79 to 0.85, indicating moderate to high discriminative ability. RF models achieved the highest AUROC (0.85, 95%CI: 0.82 – 0.88) with relatively high specificity. Boosting ensemble models demonstrated stable and well-balanced diagnostic performance. SVM and DL models showed an advantage in sensitivity. SROC curve comparisons indicated that RF and boosting ensemble models performed better in balancing sensitivity and specificity.

Conclusion

The diagnostic performance of different AI models for predicting poor functional outcomes after stroke varies across modeling approaches. Ensemble learning models, including RF and boosting-based models, showed overall advantages in specificity and in balancing sensitivity and specificity. In clinical practice, model selection should consider specific application scenarios by jointly evaluating sensitivity and specificity, and further studies are warranted to assess the clinical generalizability of AI-based prediction models.

Key words: Ischemic stroke, Functional outcome, Artificial intelligence, Machine learning, Meta-analysis

京ICP 备07035254号-20
Copyright © Chinese Journal of Cerebrovascular Diseases(Electronic Edition), All Rights Reserved.
Tel: 01082266456, 15611963912, 15611963911 E-mail: zhnxgbzzbysy@163.com
Powered by Beijing Magtech Co. Ltd